On October 6, Google launched EmbeddingGemma 2, an open-weight embedding model for developers.
Google DeepMind released EmbeddingGemma 2 on October 6, 2026, an open-weight multimodal embedding model built on the Gemma 4 architecture. Its 740 million parameters map text, code, images, video and audio into one 768-dimensional vector space, enabling cross-media search on local devices. Developers can load a 270-million-parameter text-and-code configuration or add modular vision and audio encoders; Google says the text-only model can run with about 191 MB of active RAM. The Apache 2.0 license permits commercial use, with weights available through Google and Hugging Face. The model’s 8,192-token context window is four times EmbeddingGemma’s, handling up to 5.5 minutes of audio, 29 images or 58 video frames. In Google’s reported benchmarks, it scored 78.68 on MTEB Code, up from 68.76 for its predecessor, while multilingual MTEB reached 61.36 versus 61.15.